gemini 4 argon: what it is and who can use it

as of october 2026, gemini 4 argon is google's new flagship. the google frontier model september 2026 answer is argon, not a public chatgpt style upgrade you can pick in the gemini app today. google announced it on september 30, 2026 and is shipping it first to trusted cyber defenders through the fairwind program, plus google's own teams. paid api customers and google ai ultra subscribers are next, with no date attached.

if you searched "is gemini 4 available", the accurate answer is: yes for a vetted defender cohort, not yet for everyone else.

what did google announce on september 30, 2026?

google opened a new gemini 4 generation with argon as the flagship, and it framed the release as phased on purpose.

koray kavukcuoglu, google's chief ai architect, wrote that releasing capabilities at this level "requires a phased approach". google says it is in the u.s. government's voluntary pre-release process while it hardens safeguards, and that argon already powers internal workflows for thousands of googlers.

the company highlights three product promises:

1. long horizon coding and migrations, including agents rewriting large c and c++ surfaces toward rust
2. enterprise knowledge work across finance, legal, tax, and automation style tasks
3. defensive cybersecurity, with a version for fairwind partners and internal teams that runs without the cyber guardrails applied to broader releases

argon also raises the output token limit to 1 million tokens, up from 64k on earlier gemini models, so a single trajectory can keep writing for much longer before it has to stop.

can you use gemini 4 argon yet?

you can use it if you are inside google or an approved fairwind partner. everyone else is still waiting.

the fairwind program google runs is an early access path for high priority defenders such as governments, healthcare operators, telecoms, and core technology platforms. the program page says google currently works with over 650 partners globally. siliconangle reported on september 30, 2026 that fairwind opened on september 3 with gemini 3.8 flash cyber.

fairwind partners get exclusive argon access and can pair it with codemender, google's code security agent for vulnerability research and patching. partners may use dual use tasks only for defensive and research purposes, with user level authentication, phishing resistant mfa, and no redistribution of the model.

so the search intent gemini 4 argon access resolves to an application and a waitlist, not a model picker toggle.

how does argon compare on coding and knowledge work?

google's own table puts argon ahead on several agentic and knowledge work scores, and behind on at least one important coding agent bench.

numbers below are from the deepmind gemini page and the launch post. methodology notes live on google's evals page. non gemini scores are mostly vendor reported figures that google reprinted.

that pattern matters for a gemini 4 vs opus 5.5 choice. if your work looks like long horizon software tasks of the deepswe kind, google is claiming a lead. if your work looks like terminal-bench style agent shells, our coding comparison still pointed at claude opus 5.5 before this launch, and google's own chart still shows opus ahead there.

argon also shows strong multimodal and finance or legal agent scores on the same deepmind table, including automationbench at the top of google's published set. treat every cell as a snapshot: harnesses and thinking settings differ, and the public api has not opened yet for independent rechecks.

on gemini 4 argon price, the launch post sets an introductory rate of $2 input and $10 output per million tokens, with cached input at 95% off input. see also google's gemini api pricing docs for the older public tiers while argon is gated. after the introductory period, google says argon moves to $4 input and $20 output per million tokens, matching the claude opus 5.5 api band we already tracked in ai model pricing compared.

why did google ship cyber defense first?

google is treating argon like a dual use tool that can help defenders find and patch flaws, and that can hurt if it ships wide too early.

for fairwind and internal teams, google says it will release argon without cyber guardrails so defenders can use full capability. wiz, which google owns, used argon in its scan for good work and, per the launch post, found a critical exposure in healthcare software that earlier frontier models had missed. on cwe-bench v1, argon ties gpt-6 astra at the top score in google's table.

the same week, anthropic published a separate warning about open weight cyber capability. on september 29, 2026, anthropic's frontier red team said glm-5.3 from zhipu ai can build end to end exploits at rates close to claude mythos preview. in anthropic's simulated tests, safeguards fell between 64% and 100% of the time depending on the technique. that backdrop helps explain google's fairwind staging.

google also says argon is its most resilient model yet against indirect prompt injection on gray swan's ipi benchmark, and that monitors can stop a run when chain of thought or actions drift past the user's intent. for a primer on that attack class, see what is prompt injection.

what should you do while you wait?

keep the stack you already trust, and treat argon as a scheduled change, not a reason to rebuild your workflow tonight.

practical moves:

• stay on gemini 3.1 pro for workspace heavy research until argon reaches ai ultra or the api
• keep gemini 3.8 flash for cheap background jobs; our earlier pricing notes still apply through december 31, 2026
• keep claude opus 5.5 for long agentic coding if terminal-bench style work is your daily job
• keep gpt-6 astra when computer use is the bottleneck, as covered in claude vs chatgpt vs gemini
• watch the fairwind form only if you are a defender org that meets google's criteria; consumer accounts will not skip the queue

google's internal demos are existence proofs, not your backlog. argon agents have helped migrate large codebases, including fuchsia's zircon kernel at more than 800k lines. on libgav1 they replaced 32k lines of simd oriented code with safe rust that google says runs 2.7x faster than the earlier rust port. a separate memory sweep freed over 300 tib across google data centers. those are google's numbers on google's stack.

for context windows and long document pricing quirks across labs, keep ai context window comparison nearby. argon’s 1 million token output is the new claim to watch; input window details for the public api are still thin in the launch materials.

verdict: as of october 2026, gemini 4 argon is a real frontier release with defender first access, published api intro pricing, and strong google side agent scores. it is not yet the model you should assume is live in your gemini subscription.

keep your own context while models rotate

frontier launches land every few weeks. the part that should not reset is the trail of tickets, diffs, chat logs, and browser tabs that explain why you made the last call.

remynd is a mac app built for that trail. it captures the focused window, runs ocr locally with apple vision, and keeps a searchable index on your mac. you can exclude apps and sites, and recordings are kept for 30 days by default. call transcription runs on device with mlx.

the agent window runs your own claude code or codex cli against that history with read-only access. settings accept a custom api endpoint compatible with the openai responses api. when argon finally reaches your account, the archive does not have to move with it. more on the design is in private ai on your mac and computer use agents on a mac.

download remynd for mac and keep the next model switch from wiping your working memory.